Improving Indoor Occupancy Prediction using Graph Neural Networks and Positional Encodings

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Managing residential energy efficiently without compromising comfort is challenging. The key difficulty lies in predicting occupancy in real time. Intelligent control strategies are widely used for residential energy management. As sensors become increasingly common in homes, occupancy prediction has gained significant attention as one such approach. Traditional occupancy prediction methods analyze temporal patterns within individual rooms but overlook spatial connectivity. This lack limits the accuracy of occupancy-forecasting methods. Incorporating spatial information could improve prediction accuracy and enhance the effectiveness of occupancy-based energy management. One promising approach to represent such spatial information is through the use of embeddings, which encode spatial connectivity in a data-driven vector space. Another complementary method is to represent the building layout as a graph, where nodes correspond to rooms and edges capture spatial connectivity. In this study, we evaluate multiple occupancy prediction architectures, including Convolutional Neural Networks (CNN) and Long Short-Term Memory networks (LSTM). These models are further enhanced by positional encoding (PE) and Graph Neural Networks (GNN) that explicitly encode spatial connectivity among rooms. Results indicate that spatial embeddings improve predictive performance, with graph-based methods providing the highest performance. Among all models, the hybrid GNN + PE + CNN yielded the best overall performance. However, considering practical heating control scenarios, balancing comfort, energy savings, and computational complexity, we found that LSTM + PE emerges as the most suitable choice.

This publication has 14 references indexed in Scilit: